Yuvion VL-32B Achieves Top Safety Performance, Surpassing GPT-4o

A large team of researchers have unveiled Yuvion VL, a family of multimodal large language models specifically designed to tackle adversarial content and AI safety. Unlike general-purpose models that struggle with real-world multimodal risks, Yuvion VL treats safety as an inherently adversarial problem. The 32B parameter variant, Yuvion VL-32B, is available in both instruction-tuned and reasoning-oriented versions. It is built around adversarial robustness, from data construction to training and evaluation.

The training pipeline consists of three stages: continued pretraining for cross-modal alignment of risk concepts, instruct post-training for production-grade safety tasks, and reasoning post-training to improve interpretability. A key innovation is Confuse-then-Contrast Fine-Tuning, which mines model-specific confusions and uses multi-image contrastive groups to distinguish visually similar cases with different safety implications. To evaluate performance, the authors also introduce Yuvion VL RiskEval (YVRE), a comprehensive benchmark suite for content safety, adversarial robustness, and real-world capability. Experimental results show Yuvion VL-32B outperforming comparably sized open-source models and even best closed-source commercial models like GPT-4o, while retaining strong general performance.